Multi-Layer Proximity Graphs

• Multi-layer proximity graphs are advanced structures that integrate disparate data modalities such as physical, social, and mobility to capture multifaceted relationships.
• They are constructed using spectral embeddings via normalized Laplacians and subspace merging on Grassmann manifolds, yielding robust clustering and performance metrics.
• These graphs support applications in urban analytics, community detection, and socioeconomic prediction by harmonizing multi-modal features through contrastive learning and effective fusion.

Multi-layer proximity graphs, also referred to as multi-layer or multi-modal graphs, are advanced structures used to model datasets in which relationships between entities are multifaceted, with each layer capturing a distinct notion of proximity or similarity. This abstraction allows for the integration of heterogeneous data modalities—such as geographical distance, social interactions, communication patterns, and multi-modal geotagged features—within a shared formalism, thereby enabling more comprehensive data analysis and learning tasks that exploit the richness of the underlying relationships (Dong et al., 2013, Huang et al., 2021).

1. Formal Definition and Construction

A multi-layer proximity graph consists of a vertex set $V = \{v_1, \ldots, v_n\}$ appearing in all $M$ layers of the structure. Each layer $G_i = (V, E_i, \omega_i)$ is an undirected or directed weighted graph defined over $V$, with layer-specific edge sets $E_i$ and nonnegative weight functions $\omega_i : E_i \to \mathbb{R}_+$ (Dong et al., 2013). Layers are selected to represent distinct modalities or sources of proximity:

• Physical proximity: e.g., GPS co-location.
• Interaction proximity: e.g., frequency of calls, Bluetooth scans.
• Mobility-based proximity: e.g., human trip data between neighborhoods.

Adjacency matrices $W_i \in \mathbb{R}^{n \times n}$ encode the weights for each layer, with $(W_i)_{jk} = \omega_i(\{v_j, v_k\})$. For multi-modal datasets, "point-data" (e.g., images or POIs) are associated as node features, and additional graph layers are created for each distinct edge modality, such as reciprocal distance or directional mobility (Huang et al., 2021).

2. Subspace Representation and Layer Merging

Each layer's structure can be summarized via a normalized Laplacian $L_i = D_i^{-1/2}(D_i - W_i)D_i^{-1/2}$, where $D_i$ is the degree matrix with $M$0 (Dong et al., 2013). Eigen-decomposition of $M$1 yields a $M$2-dimensional orthonormal embedding $M$3 via

$M$4

The column space $M$5 defines a point on the Grassmann manifold $M$6. To combine $M$7 such layers, subspaces are merged by finding a representative embedding $M$8 that both preserves the global multi-layer structure and minimizes the projection distance to each individual $M$9. The projection distance is defined as

$G_i = (V, E_i, \omega_i)$0

The merged subspace is obtained as the $G_i = (V, E_i, \omega_i)$1-eigenvector basis of the "modified Laplacian"

$G_i = (V, E_i, \omega_i)$2

where $G_i = (V, E_i, \omega_i)$3 controls the regularization between Laplacian structure preservation and subspace alignment (Dong et al., 2013).

3. Learning Embeddings in Multi-Layer and Multi-Modal Settings

Integration of multi-modal features and relation types requires both node and edge data harmonization. In frameworks such as the Multi-Modal Multi-Graph (M3G), each neighborhood or vertex $G_i = (V, E_i, \omega_i)$4 is represented by a learnable embedding $G_i = (V, E_i, \omega_i)$5, which is optimized so as to:

• Agree with intra-node modalities (e.g., image, text encodings from $G_i = (V, E_i, \omega_i)$6 or $G_i = (V, E_i, \omega_i)$7),
• Be close to embeddings of neighboring vertices under each edge modality (e.g., spatial distance, mobility flows).

Parameter sharing is enforced by using a single embedding $G_i = (V, E_i, \omega_i)$8 per vertex for all modalities, with no explicit GNN message passing (Huang et al., 2021).

Contrastive losses are structured as margin-based triplet objectives for each modality, e.g.,

$G_i = (V, E_i, \omega_i)$9

for image modality, and similarly for POI and various graph edges. The total loss balances the contributions from all modalities using per-modality weights $V$0, $V$1, and $V$2. Training is conducted via SGD over randomly sampled triplets (Huang et al., 2021).

4. Algorithmic Workflows

Table: Core steps in representative multi-layer proximity graph methods

Method	Construction	Embedding/Merging	Final Clustering / Use
SC-ML (Dong et al., 2013)	Build $V$3 adjacency matrices and Laplacians for $V$4 entities	Spectral embeddings $V$5 per layer, merge via Grassmann projection, solve for $V$6	$V$7-means on rows of merged $V$8
M3G (Huang et al., 2021)	Assemble node feature sets (images, POI), multiple edge adjacency matrices (distance, mobility, etc.)	Jointly optimize neighborhood embeddings with contrastive loss over all modalities	Use $V$9 for prediction, regression, or geometric clustering

In SC-ML (Spectral Clustering on Multi-Layer graphs), clustering proceeds via Laplacian-based spectral embeddings followed by $E_i$0-means in the joint embedding space. For M3G, outputs are used in tasks such as regression on socioeconomic indicators and clustering for urban analysis.

5. Quantitative Evaluation and Empirical Findings

In multi-layer spectral clustering, clustering quality is benchmarked using metrics including Purity, Normalized Mutual Information (NMI), and Rand Index (RI) (Dong et al., 2013). Multi-layer fusion achieves consistently superior results relative to single-modality or simple aggregation baselines:

• Synthetic dataset (N,R,C shapes, $E_i$1): SC-Single NMI $E_i$2, SC-Sum/KSum $E_i$3, SC-ML $E_i$4.
• NRC cellphone dataset ($E_i$5): SC-Single NMI $E_i$6, SC-Sum $E_i$7, SC-ML $E_i$8.
• Cora citation network ($E_i$9): SC-Single NMI $\omega_i : E_i \to \mathbb{R}_+$0.

For multi-modal urban neighborhood embedding, M3G embeddings outperform unimodal baselines and standard graph autoencoder methods on downstream regression tasks for census block statistics. For example, in Chicago, $\omega_i : E_i \to \mathbb{R}_+$1 for income prediction is $\omega_i : E_i \to \mathbb{R}_+$2 with M3G DIST+MOB, versus $\omega_i : E_i \to \mathbb{R}_+$3 with Urban2Vec and $\omega_i : E_i \to \mathbb{R}_+$4 with GAE (Huang et al., 2021). Combining spatial and mobility layers yields additive improvements across multiple targets.

Ablation studies confirm that spatial and mobility edge modalities contribute complementary signals to the learned embedding, while intra-node features (images, POIs) serve as semantic anchors. Qualitatively, embedding-based clusters align with semantically meaningful urban patterns, such as high-crime and low-crime neighborhoods. Embedding distance is negatively correlated with mobility volume and positively correlated with spatial distance, supporting the validity of the fused metric.

6. Applications and Significance

Multi-layer proximity graphs are foundational in problems requiring the integration of multiple sources of relational or feature-based information. They are applied in:

• Clustering, community detection: SC-ML and related approaches support unsupervised grouping when no single proximity is sufficient.
• Urban analytics: Embeddings capturing images, text, geographic, and mobility relations enable prediction of socioeconomic variables and exploration of city structure (Huang et al., 2021).
• Network analysis in multimodal social data, document corpora: Fusing citation, text, and topical similarity yields improved representations for tasks like subject-area classification (Dong et al., 2013).

A plausible implication is that as urban and relational data modalities proliferate, the formalism of multi-layer proximity graphs, together with principled embedding and fusion methods, becomes increasingly critical for robust, generalizable learning.

7. Considerations and Limitations

The effectiveness of multi-layer proximity graph methods depends on the coherence and informativeness of the layers: misaligned, noisy, or redundant modalities may degrade joint performance if not sufficiently regularized or weighted. Parameter selection, especially the regularization balancing structural faithfulness and layer agreement (e.g., $\omega_i : E_i \to \mathbb{R}_+$5 in SC-ML, or per-modality $\omega_i : E_i \to \mathbb{R}_+$6 in M3G), is critical for optimal integration.

While SC-ML employs spectral approaches and Grassmannian analysis for merging, M3G adopts a contrastive embedding regime without explicit message passing, relying on sampled triplet losses. This suggests a diversity of technique, with methodological choices informed by data type, task, and computational efficiency (Dong et al., 2013, Huang et al., 2021).